indexed

# Indexed Use the Indexed CLI as the stable Agent interface. The Skill is not the web backend and should not automate Dashboard clicks when a structured command exists. Resolve the executable before the first operation: prefer `indexed` from `PATH`; in a source checkout where it has not been linked yet, run `./bin/indexed` from the repository root. If neither resolves, report that the CLI is unavailable instead of substituting Dashboard calls. ## Operation routing - Use `indexed status` when model or storage health is unknown. - Use `indexed search "QUERY" --limit N` for retrieval. Keep visual and transcript rankings separate. - Use `indexed videos list --limit N` to resolve exact video identifiers. - Use `indexed files list`, `indexed files search`, and `indexed files scan` for configured local library roots. - Use `indexed assets libraries`, `indexed assets list`, `indexed assets search "QUERY"`, and `indexed assets scan [DIRECTORY]` for the unified asset library. It includes video clips, images, and documents inside directories the user registered, plus web-video memories captured by the Chrome extension. Local hits carry `path`; web hits carry `openUrl` and should be opened in the browser. Scanning needs a local zvec profile; report the limitation instead of scanning elsewhere. - Document search includes UTF-8 text, PDF files with a text layer, and `.doc`/`.docx`. Long documents are indexed as paragraph-aware chunks but returned as one best-passage result per file. A scanned PDF without a text layer is not searchable until an OCR workflow exists; do not report it as indexed merely because the file was discovered. - For local-file results, `indexed assets search` never returns a result whose file is missing from disk; the response counts them in `hiddenMissing`. Pass `--include-missing` to see dead rows and `--no-legacy` to search only the registered libraries. When results point at files the user deleted by hand, run `indexed assets prune --dry-run` first, report the file and row counts, then `indexed assets prune` to drop those index rows — it removes index records only, never files. - `indexed assets add DIRECTORY` registers a directory and scans it in the foreground, printing progress on stderr; pass `--no-scan` to register it and leave the work to the running server. `indexed assets scan [DIRECTORY]` (no directory: every registered root) rescans inline the same way. - Use `indexed assets add DIRECTORY` only for a directory the user named. `indexed assets remove DIRECTORY` drops records only; adding `--purge --confirm` deletes their vectors, which is irreversible and requires the user's explicit confirmation. - Use `indexed assets read FILE [--raw]` to read an editable text asset (`.md .mdx .txt .json .jsonl .csv .srt .vtt`) that already lives in a registered library, and `indexed assets write FILE --from PATH` (or `--from -` for standard input) to save it back. A save rewrites that one file in place, refreshes all of its document chunks and the excerpt its neighbouring clip quotes, and reports `vectors`, `removed`, and `previews`. PDF, Word, video and image bytes are always read-only. - Use `indexed assets model-info` before promising a scan: it reports the embedding model, its advertised context length, and the vector dimension the profile writes. - Use `indexed videos update SITE VIDEO_ID` for metadata changes requested by the user. - Use `indexed serve --open` when the user asks to see or configure the Dashboard. - Use `indexed config profiles` commands for model-space configuration. - Legacy database migration tools are not provided during this test stage. If an old index is rejected, explain how to select a fresh storage directory and rescan the registered roots; never delete the old directory automatically. Read operations and ordinary metadata updates may proceed within the user's request. Before deletion, resolve the exact `sourceSite` and `videoId`, explain that vector deletion is irreversible, obtain explicit authorization, and pass `--yes` only after that authorization. Never print credentials or infer that equal vector dimensions make two Embedding models compatible. New video capture enters through the Chrome extension after the user chooses a video. Local files are scanned only from roots already configured by the user. Do not download or silently capture unrelated content. If the TypeScript CLI command needed for a request has not yet landed during migration, report that limitation instead of fabricating a successful operation.